NVIDIA H200 vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA H200 wins 141 of 149 benchmarks, averaging 75.9% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Qwen-Image-Edit, where NVIDIA H200 leads by 255% (3.76 vs 1.06 images/min); the closest fight is TinyLlama 1.1B LoRA (1% apart); VRAM decides part of this one: NVIDIA H200 runs 152 of our 12 AI workloads while the other card runs 150, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA H200 | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 318.84 | 212.05 | +50% |
| Llama 3.1 8B tok/s | 268.31 | 135.51 | +98% |
| Qwen2.5-Coder 14B tok/s | 148.43 | 74.6 | +99% |
| Qwen3 32B tok/s | 76.58 | 34.44 | +122% |
| Llama 3.3 70B tok/s | 42.66 | 16.49 | +159% |
| Stable Diffusion XL images/min | 37.16 | 17.02 | +118% |
| Z-Image Turbo images/min | 23.175 | 8.325 | +178% |
| FLUX.1 dev images/min | 9.514 | 3.836 | +148% |
| FLUX.1 Kontext dev images/min | 4.393 | 1.693 | +159% |
| Qwen-Image-Edit images/min | 3.76 | 1.06 | +255% |
| LTX-Video (distilled) frames/s | 17.87 | 8.01 | +123% |
| Wan 2.2 5B (720p) frames/s | 1.41 | 0.49 | +188% |
| DeepSeek-R1 Distill Llama 8B tok/s | 265.25 | 135.55 | +96% |
| DeepSeek-R1 Distill 1.5B tok/s | 542.95 | 426.38 | +27% |
| DeepSeek-R1 Distill 14B tok/s | 146.82 | 74.6 | +97% |
| DeepSeek-R1 Distill 7B tok/s | 267.1 | 143.97 | +86% |
| Gemma 3 12B tok/s | 153.26 | 80.92 | +89% |
| Gemma 3 4B tok/s | 291.76 | 198.87 | +47% |
| Gemma 4 12B tok/s | 151.05 | 80.82 | +87% |
| Llama 3.2 1B tok/s | 875.53 | 618.87 | +41% |
| Llama 3.2 3B tok/s | 424.61 | 270.85 | +57% |
| Mistral 7B v0.3 tok/s | 278.5 | 144.74 | +92% |
| Mistral Small 24B tok/s | 107.84 | 48.43 | +123% |
| Phi-4 14B tok/s | 168.2 | 74.98 | +124% |
| Phi-4 Mini 3.8B tok/s | 392.38 | 228.93 | +71% |
| Qwen2.5-Coder 7B tok/s | 266.36 | 143.93 | +85% |
| Qwen3 0.6B tok/s | 718.67 | 655.75 | +10% |
| Qwen3 1.7B tok/s | 563.52 | 405.52 | +39% |
| Qwen3 14B tok/s | 154.51 | 75.12 | +106% |
| Qwen3 30B A3B tok/s | 292.11 | 213.66 | +37% |
| Qwen3 8B tok/s | 247.98 | 130.27 | +90% |
| SmolLM3 3B tok/s | 402.24 | 272.88 | +47% |
| TRELLIS Image-to-3D assets/hour | 818.8 | 776.1 | +6% |
| Codestral 22B tok/s | 111.34 | 50.06 | +122% |
| DeepSeek-R1 Distill 32B tok/s | 75.8 | 34.6 | +119% |
| Devstral Small 24B tok/s | 109.05 | 48.43 | +125% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 76.7 | 33.08 | +132% |
| Dolphin Mistral 24B Venice tok/s | 109.14 | 48.43 | +125% |
| Dolphin X1 8B tok/s | 268.36 | 135.48 | +98% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 267.84 | 135.55 | +98% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 109.06 | 48.4 | +125% |
| Gemma 3 27B tok/s | 84.75 | 38.48 | +120% |
| Qwen2.5-Coder 32B tok/s | 75.77 | 34.59 | +119% |
| Qwen3-Coder 30B A3B tok/s | 297.9 | 219.04 | +36% |
| QwQ 32B tok/s | 75.77 | 34.59 | +119% |
| StarCoder2 15B tok/s | 136.29 | 66.88 | +104% |
| FLUX.1 Schnell images/min | 60.51 | 26.24 | +131% |
| Z-Image Turbo (1024px) images/min | 40.74 | 18.2 | +124% |
| BiRefNet images/min | 1457.73 | 863.81 | +69% |
| Depth Anything V2 Large images/min | 1040.81 | 959.33 | +8% |
| Depth Anything V2 Small images/min | 1198.58 | 970.61 | +23% |
| SAM ViT-Base images/min | 1469.34 | 998.79 | +47% |
| SAM ViT-Huge images/min | 325.19 | 238.08 | +37% |
| Swin2SR 4x Upscaler images/min | 25.85 | 34.14 | -24% |
| Qwen2.5 1.5B LoRA train tok/s | 14931.7 | 12120.1 | +23% |
| Qwen2.5 7B LoRA train tok/s | 8855.4 | 3736 | +137% |
| SmolLM2 1.7B LoRA train tok/s | 17650.4 | 12052.1 | +46% |
| TinyLlama 1.1B LoRA train tok/s | 16222.4 | 16408.1 | -1% |
| Qwen2.5 1.5B served serve tok/s | 7541.6 | 4686.8 | +61% |
| Qwen2.5 7B served serve tok/s | 4394.8 | 1406.8 | +212% |
| SmolLM2 1.7B served serve tok/s | 7107.8 | 3525.9 | +102% |
| TinyLlama 1.1B served serve tok/s | 9137.7 | 6278.4 | +46% |
| Kokoro TTS 82M x realtime | 170.82 | 243.99 | -30% |
| MusicGen Small x realtime | 1.87 | 2.43 | -23% |
| Whisper large-v3 x realtime | 163.78 | 193.2 | -15% |
| AI21-Jamba-Reasoning-3B tok/s | 370.66 | 258.41 | +43% |
| Olmo-3.1-32B-Think tok/s | 78.46 | 34.11 | +130% |
| Codestral 22B (Q3_K_M) tok/s | 88.03 | 59.53 | +48% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 109.2 | 48.43 | +125% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 267.82 | 134.78 | +99% |
| DeepSeek-Coder-V2-Lite tok/s | 312.1 | 257.64 | +21% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 250.61 | 130.23 | +92% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 42.76 | 16.49 | +159% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 119.66 | 85.85 | +39% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 75.71 | 34.43 | +120% |
| dolphin-2.9-llama3-8b tok/s | 267.96 | 134.94 | +99% |
| Dolphin X1 Trinity Nano 6B tok/s | 259.81 | 246.45 | +5% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 148.74 | 74.63 | +99% |
| gemma-2-2b-it-abliterated tok/s | 397.43 | 278.14 | +43% |
| gemma-2-2b tok/s | 398.48 | 278.1 | +43% |
| gemma-2-9b tok/s | 180.47 | 88.87 | +103% |
| Gemma 3 12B (Q3_K_M) tok/s | 127.04 | 93.05 | +37% |
| gemma-3-1b tok/s | 513.58 | 426.49 | +20% |
| gemma-3-270m tok/s | 967.35 | 822.52 | +18% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 171.45 | 146.51 | +17% |
| GLM-4.7-Flash tok/s | 188.57 | 157.89 | +19% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 250.88 | 130.25 | +93% |
| gpt-oss-20b tok/s | 354.23 | 233.82 | +51% |
| Hermes-3-Llama-3.2-3B tok/s | 430.79 | 265.87 | +62% |
| Hermes-4-70B tok/s | 42.77 | 16.45 | +160% |
| SmolLM3-3B tok/s | 404.27 | 272.96 | +48% |
| Qwen3-Coder-Next-abliterated tok/s | 197.14 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 236.31 | 173.06 | +37% |
| L3-8B-Stheno-v3.2 tok/s | 267.92 | 135.53 | +98% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 946.53 | 661.54 | +43% |
| LFM2.5-8B-A1B tok/s | 583.96 | 394.48 | +48% |
| Llama-2-7B tok/s | 290.81 | 150.35 | +93% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 429.78 | 271.01 | +59% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 42.75 | 16.49 | +159% |
| Meta-Llama-3.1-70B tok/s | 42.72 | 16.49 | +159% |
| Meta-Llama-3.1-8B tok/s | 262.93 | 135.46 | +94% |
| Phi-4-mini tok/s | 395.3 | 229.28 | +72% |
| Mistral-7B-Instruct-v0.1 tok/s | 282.64 | 144.82 | +95% |
| Mistral-7B-Instruct-v0.2 tok/s | 282.32 | 144.82 | +95% |
| Mistral-7B-Instruct-v0.3 tok/s | 282.39 | 144.68 | +95% |
| Mistral-Nemo-Instruct-2407 tok/s | 181.3 | 89.72 | +102% |
| Mistral Small 24B (Q3_K_M) tok/s | 85.1 | 58.21 | +46% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 181.53 | 88.85 | +104% |
| Nemotron-3-Nano-30B-A3B tok/s | 327.48 | 190.96 | +71% |
| Hermes-4-14B tok/s | 156.26 | 75.1 | +108% |
| Ornith-1.0-35B tok/s | 208.51 | 157.39 | +32% |
| Ornith-1.0-9B tok/s | 222.94 | 114.57 | +95% |
| phi-2 tok/s | 347.08 | 283.26 | +23% |
| Phi-3.5-mini tok/s | 348.14 | 229.06 | +52% |
| Phi-4 14B (Q3_K_M) tok/s | 139.97 | 89.18 | +57% |
| Qwen-AgentWorld-35B-A3B tok/s | 225.31 | 156.98 | +44% |
| Qwen3-0.6B tok/s | 724.67 | 656.26 | +10% |
| Qwen3-1.7B tok/s | 568.85 | 405.3 | +40% |
| Qwen3-14B tok/s | 156.3 | 75.01 | +108% |
| Qwen3-30B-A3B tok/s | 293.47 | 214.05 | +37% |
| Qwen3-4B-Instruct-2507 tok/s | 318.67 | 210.24 | +52% |
| Qwen3-8B tok/s | 251.04 | 130.28 | +93% |
| Qwen3-Coder-Next tok/s | 195.21 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 194.82 | 0 | n/a |
| Qwen1.5-0.5B tok/s | 855.22 | 774.78 | +10% |
| Qwen2-1.5B tok/s | 546.47 | 416.57 | +31% |
| Qwen2.5-0.5B tok/s | 914.58 | 738.06 | +24% |
| Qwen2.5-1.5B tok/s | 549.36 | 427.23 | +29% |
| Uncensored tok/s | 148.7 | 74.31 | +100% |
| Qwen2.5-14B tok/s | 148.76 | 74.62 | +99% |
| Qwen2.5-32B tok/s | 75.76 | 34.59 | +119% |
| Qwen2.5-3B tok/s | 400.48 | 270.04 | +48% |
| Qwen2.5-7B tok/s | 265.15 | 143.94 | +84% |
| Qwen2.5-Coder-0.5B tok/s | 920.36 | 715.81 | +29% |
| Qwen2.5-Coder-1.5B tok/s | 545.54 | 427.12 | +28% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 148.48 | 74.63 | +99% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 58.87 | 41.06 | +43% |
| Qwen2.5-Coder-3B tok/s | 401.11 | 269.9 | +49% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 270.92 | 142.97 | +89% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 246.35 | 224.95 | +10% |
| Qwen3-4B-Instruct-2507 tok/s | 318.27 | 212.05 | +50% |
| Qwen3-4B-Thinking-2507 tok/s | 318.94 | 211.88 | +51% |
| Qwen3-Coder-Next tok/s | 196.65 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 201.12 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 192.52 | 0 | n/a |
| SmolLM2-135M tok/s | 905.38 | 916.56 | -1% |
| Cydonia-24B-v4.3 tok/s | 109.2 | 48.43 | +125% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA H200 is faster on 15 of 17; NVIDIA L40S on 2.
| Workflow | NVIDIA H200 | NVIDIA L40S | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 6 s | 6 s | NVIDIA L40S 1.02x faster | $0.006 vs $0.001 |
| 30-minute podcast pass | 46 s | 75 s | NVIDIA H200 1.65x faster | $0.046 vs $0.017 |
| 500-image masking run | 1.7 min | 2.2 min | NVIDIA H200 1.29x faster | $0.100 vs $0.028 |
| 20-asset 3D game kit | 2.4 min | 2.9 min | NVIDIA H200 1.21x faster | $0.145 vs $0.039 |
| 6-panel comic page | 3.2 min | 6.5 min | NVIDIA H200 2.03x faster | $0.192 vs $0.086 |
| 60-second AI short film | 3.9 min | 4.9 min | NVIDIA H200 1.27x faster | $0.233 vs $0.065 |
| Character sheet, 12 poses | 3.9 min | 8.4 min | NVIDIA H200 2.14x faster | $0.234 vs $0.110 |
| 24-frame storyboard | 3.9 min | 4.1 min | NVIDIA H200 1.04x faster | $0.234 vs $0.053 |
| 60-second AI short film, narrated | 4.1 min | 5.1 min | NVIDIA H200 1.25x faster | $0.243 vs $0.067 |
| Full codebase review | 6.7 min | 13.5 min | NVIDIA H200 2.00x faster | $0.403 vs $0.177 |
| 200-product catalogue cutout | 8 min | 6.2 min | NVIDIA L40S 1.29x faster | $0.479 vs $0.082 |
| 10 short social clips | 10 min | 19.1 min | NVIDIA H200 1.91x faster | $0.598 vs $0.251 |
| 20 long-form articles | 10.9 min | 28.5 min | NVIDIA H200 2.60x faster | $0.655 vs $0.375 |
| 40-product photo shoot | 11.2 min | 26.7 min | NVIDIA H200 2.39x faster | $0.668 vs $0.352 |
| 40-product shoot, start to finish | 12.9 min | 28 min | NVIDIA H200 2.18x faster | $0.770 vs $0.369 |
| 100-photo restoration batch | 23.3 min | 59.6 min | NVIDIA H200 2.55x faster | $1.396 vs $0.784 |
| 100-photo restore and enlarge | 27.2 min | 62.5 min | NVIDIA H200 2.30x faster | $1.629 vs $0.823 |
Renting by the hour, NVIDIA L40S finishes 17 of 17 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA H200 | $3.590 |
| NVIDIA L40S | $0.790 |
NVIDIA L40S is 4.54x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA H200 | NVIDIA L40S | |
|---|---|---|
| VRAM | 141GB | 48GB |
| Architecture | Hopper | Ada Lovelace |
| Memory bandwidth | 4800 GB/s | 864 GB/s |
| Boost clock | 1,980 MHz | 2,520 MHz |
| TDP | 700 W | 350 W |
| Launch MSRP | $31,000 | $7,500 |
| Release | 2024-03-18 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA H200 or NVIDIA L40S?
- NVIDIA H200 performs better for ai & machine learning, winning 141 of 149 benchmarks in our suite with an average 75.9% advantage.
- What are the main hardware differences between NVIDIA H200 and NVIDIA L40S?
- NVIDIA H200 has 141GB VRAM and a 700W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA H200 and NVIDIA L40S?
- For AI, yes, NVIDIA H200 fits 6 more of our 12 workloads than NVIDIA L40S. A model that exceeds VRAM doesn't run slower, it doesn't run at all, so the card that fits the model wins that workload outright.
- Where is the biggest performance difference between NVIDIA H200 and NVIDIA L40S?
- Qwen-Image-Edit: NVIDIA H200 leads by roughly 255% (3.76 vs 1.06 images/min) in our testing.
NVIDIA H200 full review · NVIDIA L40S full review · All AI & Machine Learning rankings